舞台场景下结合姿态估计的演员检测方法  

Actor Detection Combined with Pose Estimation in Stage Scenes

作　　者：胡锆 董天阳[1] 方思琦 江一鸣 HU Gao;DONG Tian-yang;FANG Si-qi;JIANG Yi-ming(College of Computer Science and Technology,Zhejiang University of Technology,Hangzhou 310023,China)

机构地区：[1]浙江工业大学计算机科学与技术学院,杭州310023

出　　处：《小型微型计算机系统》2023年第11期2603-2609,共7页Journal of Chinese Computer Systems

基　　金：国家重点研发计划项目(2018YFB1404100)资助;国家自然科学基金项目(62072405)资助;浙江省自然科学基金项目(LGF20F020017)资助。

摘　　要：演员检测是将混合现实(MR)技术应用于戏剧表演的关键技术之一.然而,现有的行人检测方法无法在复杂舞台场景中进行有效的演员检测.因此为了提高演员检测的准确率,本文提出了一种用于复杂舞台场景的演员检测网络,将演员的关键点语义信息引入演员检测中.首先使用姿态估计子网络预测每个演员的关键点,然后根据演员的关键点生成包含语义信息的热力图.接着采用注意力机制将语义信息与原始图像融合,融合后的特征图送入检测网络进行进一步检测.同时为了提高关键点预测的准确率,改进了姿态估计子网络,将ResNet翻转拼接,堆叠成两级沙漏结构网络.在舞台演员数据集上进行对比检测,实验结果表明,引入关键点语义信息后的演员检测网络在MR,AP和AR上分别达到了14.37%,68.4%,73.3%要明显优于当前流行的其他方法.Actor detection is one of the key technologies in the application of mixed reality(MR)technology to theatrical performances.However,the existing pedestrian detection methods can not carry out effective actor detection in the stage scenes under complex illumination conditions.In order to enhance the accuracy of actor detection,this paper presents a novel actor-detection network for complex illumination.This actor-detection network introduced the semantic information of actor's pose into pedestrian detector.This method uses a sub-network to detect the keypoints of each pedestrian,and then generates a heatmap containing semantic information based on the keypoints of actors.In addition,this method adopts the attention mechanism to fuse the semantic information with the original image,and the fused feature map is past through the detection network for further detection.In order to improve the accuracy of detecting keypoints,this method flips and splices the ResNet,then stacks them into a two-stage hourglass structure network.Comparative detection is performed on the stage actor dataset.The experimental results show that the actor detection network after introducing the semantic information of key points achieves 14.37%,68.4%and 73.3%on MR,AP and AR,respectively,which is significantly better than other popular methods.

关 键 词：演员检测 姿态估计 关键点 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]